Read a string in a cell as a string, in the binary dictionary too

[ww-mw]

{"a"} in a compressed-binary .sldd displayed as {<1x1 string>} — the summary form, so the text was not on screen at all and the cell offered no editor. The text dictionary showed {"a"} the whole time, which is how it was reported: the same value read two ways depending on which format it had been saved in.

A string is not a JSON value in the XML channel, it is an object with a saveobj payload, and BinarySlddParser nests one in three places — parseEntryValue (an entry's own value), parsePropContent (a struct field or an object property), parseCellElement (a cell element). The first two tested Class="string"; the third never did, so the element fell through to the generic nested-object tail and became an object of class string whose text sat in the saveobj bag, where no formatter looks.

The display was the visible tip. Nothing decoded that bag back, so rebuilding the chunk wrote the element out as an empty <Element Class="string"> and the text was gone from the file — silent data loss on any save of a dictionary that held one, with no edit to it required.

Fixed by lifting string out of the object tail in parseCellElement, the same two lines its two siblings already had.

How the shape was settled

test/parity/matlab/probe_cell_string.m, against R2027a: a cell element holding an object is a classless <Element> wrapping the object's own <Element Class="...">, one level deeper than it looks — so the tail is right for every other class and the fix had to be a lift, not a replacement. Its two fixtures (cellstr_text.sldd, cellstr_binary.sldd) are one dictionary in both flavours, which lets the channels be asserted against each other rather than each against its own literal.

Tests

• binaryWriteBackGate.test.ts — reads MATLAB's own bytes, compares both channels, and pins that an untouched rebuild keeps all six saveobj payloads and reopens as the values MATLAB wrote; two entries are asserted byte-identical to MATLAB's chunk.
• binarySlddValues.test.ts — the reader-level regression, failing before the fix.
• cellElementShape.test.ts — three assertions that had pinned {<1x1 string>} as a known limitation of the binary channel, blamed on a .mat MCOS payload, now assert both channels agree. An .sldd carries no MCOS blob and the write side was already byte-correct, so the limitation bullet they cited now says so.

Full npm run verify green: 4868 tests.

Two neighbours recorded, neither touched

• We write Dimension="1*1" on a 1x1 cell's <P Name="Value"> where MATLAB omits the attribute. Save churn, not data loss; it round-trips, and it predates this defect.
• MATLAB build 27.1.0.3393633 no longer emits _fields in a text dictionary, and StructNode.parse builds a scalar struct's field rows from _fields alone — so a struct in a current-MATLAB text .sldd shows no field row at all. Wider blast radius than this defect and not fixed here; measured with a nine-struct probe and written up in DESIGN.md.